Researchers from Waabi and the University of Toronto have developed LabelFormer, a transformer-based AI model that efficiently refines object trajectories for auto-labelling. This technique improves the accuracy of bounding boxes by utilizing the entire time context and outperforms window-based approaches in terms of computing efficiency. The comprehensive experimental assessment demonstrates the effectiveness of LabelFormer in auto-labelling larger datasets for training item detectors.
Researchers from Waabi and the University of Toronto Introduce LabelFormer: An Efficient Transformer-Based AI Model to Refine Object Trajectories for Auto-Labelling
Modern self-driving systems require large-scale datasets to train object detectors for recognizing traffic participants. However, manually annotating these datasets can be expensive and time-consuming. Auto-labeling methods, which automatically produce sensor data labels, have recently gained attention as a more cost-effective alternative.
LabelFormer, developed by researchers from Waabi and the University of Toronto, is a straightforward and effective trajectory refining technique that produces more precise bounding boxes. It leverages the temporal context of the complete object trajectory to overcome challenges such as object occlusions, sparsity of observations, and various object sizes and motion patterns.
LabelFormer eliminates unnecessary computations by refining the entire trajectory in a single shot, making it more computationally efficient. It also outperforms window-based approaches and offers a distinct advantage over human annotation in terms of computing efficiency.
The researchers created a transformer-based architecture using self-attention blocks to capture dependencies over time. This approach significantly improves the performance of auto-labeling and allows for the training of more accurate object detectors.
LabelFormer has been extensively evaluated on highway and urban datasets, demonstrating its speed and performance superiority over window-based methods. It can auto-label larger datasets for training downstream object detectors, resulting in more accurate detections compared to human annotation alone or other auto-labeling techniques.
If you’re looking to evolve your company with AI and stay competitive, consider using LabelFormer to refine object trajectories and improve auto-labeling. Contact us at hello@itinai.com for AI KPI management advice and explore solutions at itinai.com to redefine your sales processes and customer engagement with AI.
Spotlight on a Practical AI Solution – AI Sales Bot
Discover how AI can redefine your sales processes and customer engagement with the AI Sales Bot from itinai.com/aisalesbot. This AI-powered bot is designed to automate customer engagement 24/7 and manage interactions across all stages of the customer journey.
Explore the solutions at itinai.com to harness the power of AI in your company.
